demo/robustReg_gbmt.r

In gbm-developers/gbm3: Generalized Boosted Regression Models

library(MASS)

set.seed(20090415)


x <- mvrnorm(100, mu=rep(0,5) , Sigma=diag(rep(1,5)))
r <- rnorm(100)
r <- ifelse(runif(100) < 0.25, r*4, r)
y <- apply(x, 1, sum) + r

d <- data.frame(y=y, x)
train_params <- training_params(num_trees = 2000, 
                                interaction_depth = 1, 
                                bag_fraction = 0.5,
                                num_train = nrow(d), 
                                id = seq_len(nrow(d)), 
                                num_features = ncol(x),
                                shrinkage = 0.01)

gmod <- gbmt(y ~ ., data=d, distribution=gbm_dist("Gaussian"),
             train_params = train_params, cv_folds=5, is_verbose = FALSE)
tmod4 <- gbmt(y ~ ., data=d, distribution=gbm_dist("TDist"), # defaults to 4 df
             train_params = train_params, cv_folds=5, is_verbose = FALSE)
tmod6 <- gbmt(y ~ ., data=d, distribution=gbm_dist(name="TDist", df=6),
              train_params = train_params,cv_folds=5, is_verbose = FALSE)
tmod100 <- gbmt(y ~ ., data=d, distribution=gbm_dist(name="TDist", df=100),
               train_params = train_params, cv_folds=5, is_verbose = FALSE)

par.old <- par(mfrow=c(2,2))
gbest <- gbmt_performance(gmod , method="cv")
t4best <- gbmt_performance(tmod4 , method="cv")
t6best <- gbmt_performance(tmod6 , method="cv")
t100best <- gbmt_performance(tmod100 , method="cv")
par(par.old)